What's new

Particle-1’s temporary checks caught a 2048 merge bug, inspected a maze’s structure and tested Matrix Rain’s pause behavior.

OUTPUT

particle-1

OUTPUT

particle-1

OUTPUT

particle-1

GPT-6.1 Terra paired browser checks with targeted game tests. Compare its movement and collision checks with GPT-6.1 Sol’s earlier debugging traces.

OUTPUT

gpt-6.1-terra

OUTPUT

gpt-6.1-sol

OUTPUT

gpt-6.1-terra

OUTPUT

gpt-6.1-sol

A coding agent says it has tested the app in a browser. You open it anyway, click the buttons, change the inputs, and check it yourself. You delegated verification, but still cannot tell which parts of the result you can rely on.

Kimi K3.1 Pro combined short browser-tool traces with scripted game checks, including a playable route through Platformer Dave.

OUTPUT

kimi-k3.1-pro

OUTPUT

kimi-k3.1-pro

OUTPUT

kimi-k3.1-pro

OUTPUT

kimi-k3.1-pro

Mercury 3 made game and UI state readable as text, then used that feedback to revise its builds.

OUTPUT

mercury-3

OUTPUT

mercury-3

OUTPUT

mercury-3

MiniMax revisited every build after editing, but repeated revisions sometimes consumed the full time budget.

OUTPUT

minimax-m3.1-flash

OUTPUT

minimax-m3.1-flash

OUTPUT

minimax-m3.1-flash

Keep interesting builds and models close at hand, return to them later, and open your saved models side by side. No account needed.

Nemotron’s cube shows a concrete repair sequence; Muse Spark Open’s saved runs contain fewer browser checks.

OUTPUT

nemotron-3.5-super-vl

OUTPUT

muse-spark-open

OUTPUT

nemotron-3.5-super-vl

173 successful model/prompt builds published in autonomous sessions with browser tools. See each run for its output, transcript, runtime and cost.

OUTPUT · TIMED OUT

qwen3-vl-30b-a3b-instruct

OUTPUT

qwen-2.5-72b-instruct

OUTPUT

qwen3.5-35b-a3b

OUTPUT

qwen3-coder-next

3 successful model/prompt builds published in autonomous sessions with browser tools. See each run for its output, transcript, runtime and cost.

OUTPUT

gemini-4-argon

OUTPUT

gemini-4-argon

OUTPUT

gemini-4-argon

Ling repeatedly revised after browser checks; Muse Wonder exercised controls across every prompt.

OUTPUT

ling-3.1-flash

OUTPUT

muse-wonder

OUTPUT

ling-3.1-flash

52 successful model/prompt builds published in autonomous sessions with browser tools. See each run for its output, transcript, runtime and cost.

OUTPUT

mimo-v2.6-flash

OUTPUT

gemma-4-26b-a4b-it

OUTPUT

minimax-m3

OUTPUT

gemma-4-26b-a4b-it

We have published 496 successful browser-agent builds across 18 models. These outputs were generated in earlier runs and saved in Cloudflare; this release brings them into the gallery without rerunning the models.

OUTPUT

gpt-oss-120b

OUTPUT

qwen3-8b

OUTPUT

nemotron-3-nano-30b-a3b

OUTPUT · TIMED OUT

laguna-s-2.1

174 successful model/prompt builds published in autonomous sessions with browser tools. See each run for its output, transcript, runtime and cost.

OUTPUT

claude-haiku-5.5

OUTPUT

qwen3.8-flash

OUTPUT

qwen3.6-35b-a3b

OUTPUT

gpt-5-mini

139 successful model/prompt builds published in autonomous sessions with browser tools. See each run for its output, transcript, runtime and cost.

OUTPUT

qwen3.8-27b

OUTPUT

mimo-v2.6-pro

OUTPUT

gemini-3.5-flash-lite

OUTPUT

step-5-preview

139 successful model/prompt builds published in autonomous sessions with browser tools. See each run for its output, transcript, runtime and cost.

OUTPUT

gpt-5-nano

OUTPUT

gpt-6-luna

OUTPUT

glm-5.3-flash

OUTPUT

gpt-5-nano

140 successful model/prompt builds published in autonomous sessions with browser tools. See each run for its output, transcript, runtime and cost.

OUTPUT

gpt-5.6-terra

OUTPUT

gemini-3.1-pro-preview

OUTPUT

grok-4.7

OUTPUT

qwen3.8-max-0902

DeepSeek V4.1 Flash has 34 complete outputs in this refresh across the current 35 prompts, with HTML, previews, transcripts and OpenRouter-reported usage available on each run page.

OUTPUT

deepseek-v4.1-flash

OUTPUT

deepseek-v4.1-flash

OUTPUT

deepseek-v4.1-flash

OUTPUT

deepseek-v4.1-flash

Gemini 3.8 Flash and Muse Spark 1.3 are now available across the current 35 prompts. This batch produced 69 complete outputs: 34 from Gemini and all 35 from Muse, with HTML, previews, transcripts and OpenRouter-reported usage on each run page.

OUTPUT

gemini-3.8-flash

OUTPUT

muse-spark-1.3

OUTPUT

gemini-3.8-flash

OUTPUT

muse-spark-1.3

This refresh publishes 322 complete results across ten models, each with HTML, previews, a transcript and OpenRouter-reported usage. The models ran the current 35 prompts with the pi harness and browser tools available.

OUTPUT · STOPPED BY AN API ERROR

gemma-4-26b-a4b-it

OUTPUT

qwen-2.5-7b-instruct

OUTPUT

nova-lite-v1

OUTPUT

qwen3.5-flash-02-23

This refresh publishes 340 complete results across ten models, each with HTML, previews, a transcript and OpenRouter-reported usage. The models ran the current 35 prompts with the pi harness and browser tools available.

OUTPUT

ministral-14b-2512

OUTPUT

step-3.5-flash

OUTPUT

gemma-4-31b-it

OUTPUT · STOPPED BY AN API ERROR

qwen3-235b-a22b-2507

Fresh runs of 35 prompts are now published for five models:

OUTPUT

mistral-nemo

OUTPUT

ling-3.0-flash

OUTPUT

gpt-oss-20b

OUTPUT

llama-3.1-8b-instruct

Mistral Large 4 joins the gallery with runnable outputs for all 35 current prompts.

OUTPUT · STOPPED BY AN API ERROR

mistral-large-4-0

OUTPUT · STOPPED BY AN API ERROR

mistral-large-4-0

OUTPUT · STOPPED BY AN API ERROR

mistral-large-4-0

OUTPUT · STOPPED BY AN API ERROR

mistral-large-4-0

A fresh run on our rewritten prompts: 33 complete builds with browser iteration and new previews.

OUTPUT

deepseek-v4.1-flash

OUTPUT

deepseek-v4.1-flash

OUTPUT

deepseek-v4.1-flash

OUTPUT

deepseek-v4.1-flash

Claude Sonnet 5.5 produced HTML for all 35 prompts, checking its work in the browser on every run.

OUTPUT

claude-sonnet-5.5

OUTPUT

claude-sonnet-5.5

OUTPUT

claude-sonnet-5.5

OUTPUT

claude-sonnet-5.5

GPT-6.1 Sol produced HTML for all 35 prompts, using browser inspection and iteration in our new pipeline.

OUTPUT

gpt-6.1-sol

OUTPUT

gpt-6.1-sol

OUTPUT

gpt-6.1-sol

OUTPUT

gpt-6.1-sol

We added the smaller Qwen3.8-27B, so we put it straight up against the 2.4-trillion-parameter Qwen3.8-2.4T-A95B already in the gallery — same family, roughly two orders of magnitude apart in size. Both are reliable builders (34 of 35 and 35 of 35 shipped), so the story is entirely in the quality.

OUTPUT · TIMED OUT

qwen3.8-27b

OUTPUT

qwen3.8-2.4t-a95b

OUTPUT

qwen3.8-27b

OUTPUT

qwen3.8-2.4t-a95b

xAI's Grok 4.6 ran the full matrix clean — 35 of 35 builds shipped — with a clear strength in generative canvas work.

OUTPUT

grok-4.6

OUTPUT

grok-4.6

OUTPUT

grok-4.6

OUTPUT

grok-4.6

Meta shipped two new Muse models, so we ran both through the same 35 one-shot prompts. Neither failed to ship a single build — 35 of 35 runnable each — so the gap between them is entirely about quality, not reliability.

OUTPUT

muse-spark-1.2

OUTPUT

muse-glimmer-30b

OUTPUT

muse-spark-1.2

OUTPUT

muse-glimmer-30b

Alibaba's 2.4-trillion-parameter Qwen3.8-2.4T-A95B turned in the most consistent run of the batch — 35 of 35 builds shipped, zero blank or crashed, and the tightest quality spread in the set.

OUTPUT

qwen3.8-2.4t-a95b

OUTPUT

qwen3.8-2.4t-a95b

OUTPUT

qwen3.8-2.4t-a95b

OUTPUT

qwen3.8-2.4t-a95b

Two new ByteDance Seed models went through the matrix together — both 35 of 35 on build reliability, but with opposite temperaments.

OUTPUT

seed-2.0-code

OUTPUT

seed-2-1-turbo

OUTPUT

seed-2.0-code

OUTPUT

seed-2-1-turbo

Alibaba's new flagship Qwen3.8-Max ran all 35 one-shot prompts with zero build failures — a clean sweep of the matrix, and a solid mid-pack showing on quality.

OUTPUT

qwen3.8-max

OUTPUT

qwen3.8-max

OUTPUT

qwen3.8-max

OUTPUT

qwen3.8-max

A new side-by-side grid: pick up to 10 models and see how each one built every prompt, in one scrollable table.

We ran Moonshot's new Kimi K3 and Claude Opus 4.8 through the same 34 one-shot prompts, same harness, and they finish within a hair of each other — K3 at 3.24 average, Opus at 3.12. The interesting part is where each one misses.

OUTPUT

kimi-k3

OUTPUT

claude-opus-4.8

OUTPUT

kimi-k3

OUTPUT

claude-opus-4.8

Moonshot shipped Kimi K3, a million-token model, so we added it to the matrix and ran all 34 one-shot prompts against it — same harness, same single-shot rules as every other build in the gallery.

OUTPUT

kimi-k3

OUTPUT

kimi-k3

OUTPUT

kimi-k3

OUTPUT

kimi-k3

Five more models joined the matrix this week, and we re-ran every prompt against them. The alien-shooter build is a good stress test: it needs a game loop, collision, and input handling all in one file, so the gap between models shows up fast.

OUTPUT

deepseek-v4-pro

OUTPUT

kimi-k2.7-code

OUTPUT

glm-5.2

OUTPUT

minimax-m2.7

Every build now carries a two-tier grade. Tier 1 measures the artifact deterministically — how much of the screen moves, whether it responds to input, console and JS errors — and Tier 2 hands those measurements plus a capture to a vision model for the final verdict.

OUTPUT

claude-opus-4.8

OUTPUT

deepseek-v4-pro

OUTPUT

claude-opus-4.8

OUTPUT

deepseek-v4-flash

A real-time WebGL fluid sim in one file — the hardest prompt in the set so far.

OUTPUT

deepseek-v4-pro

OUTPUT

glm-5

OUTPUT

minimax-m2.7

OUTPUT

claude-opus-4.8